Transcript IB 372 Lab 1: Introduction to Statistics

IB 372
Lab 1: Introduction to Statistics
Fall 2010
Thanks to Steve Paton - Smithsonian Tropical Research
Institute for providing original spanish version of this
file…
What are statistics?
• Statistics are numbers used to:
Describe and draw conclusions about DATA
• These are called descriptive (or “univariate”) and
inferential (or “analytical”) statistics, respectively.
Statistic vs. Parameter
Formally (and somewhat confusingly):
• A statistic is a measure of some attribute of a
sample.
• Whereas, a parameter is the real and unique
measure of the attribute for the whole
population.
• Usually, the population is too big to measure, so
in practice, statistics represent parameters.
(thus, even “descriptive” stats are usually inferential too)
Variables
• A variable is anything we can measure/observe
• Three types:
– Continuous: values span an uninterrupted range (e.g. height)
– Discrete: only certain fixed values are possible (e.g. counts)
– Categorical: values are qualitatively assigned (e.g. low/med/hi)
• Dependence in variables:
“Dependent variables depend on independent ones”
– Independent variable may or may not be prescribed
experimentally
– Determining dependence is often not trivial!
Descriptive Statistics
Descriptive statistics
Techniques to summarize data
Numerical
– Mean
– Variance
• Standard deviation
• Standard error
–
–
–
–
Median
Mode
Skew
etc.
Graphical
–
–
–
–
Histogram
Boxplot
Scatterplot
etc.
The Mean:
Most important measure of “central tendency”
Population Mean
N
m=
X
S i
i=1
N
The Mean:
Most important measure of “central tendency”
Sample Mean
n
X=
X
S i
i=1
n
Additional central tendency measures
M = X(n+1)/2
Median: the 50th percentile
M=
(n is odd)
Xn/2 + X(n/2)+1
2
(n is even)
Mode: the most common value
1, 1, 2, 4, 6, 6, 6, 7, 7, 7, 7, 8, 8, 9, 9, 10, 12, 15
Which to use: mean, median or mode?
Variance:
Most important measure of “dispersion”
Population Variance
2
s
=
S (Xi N
2
µ)
Variance:
Most important measure of “dispersion”
Sample Variance
2
s
=
S (Xi -
2
X)
n-1
From now on, we’ll ignore sample vs. population. But remember:
We are almost always interested in the population, but can measure only a sample.
Additional dispersion measures
s =√
s2
Standard deviation:
average distance from the mean
Standard error:
the accuracy of our estimate of the
population mean
(duh!)
s
SE =
√n
Bigger sample size (n)
 smaller error
Range: total span of the data (Xmax - Xmin)
Additional dispersion measures
Coefficient of Determination:
Standard deviation scaled to data
Example:
1.2
1.4
1.6
1.8
2.0
2.2
2.4
X = 1.8kg
s = 0.43kg
V = 24%
vs.
1200
1400
1600
1800
2000
2200
2400
s
V=
X
X = 1800kg
s = 432kg
V = 24%
“Graphical Statistics”
The Friendly Histogram
• Histograms represent the distribution of data
• They allow you to visualize the mean, median,
mode, variance, and skew at once!
Constructing a Histogram is Easy
Histogram of X
X (data)
Frequency
(count)
7.4
7.6
8.4
8.9
10.0
10.3
11.5
11.5
12.0
12.3
3
2
1
0
6
8
10
Value
12
14
Interpreting Histograms
Mean?
Median?
Mode?
Standard
deviation?
• Variance?
• Skew?
40
Frequency
•
•
•
•
20
(which way does the
“tail” point?)
• Shape?
0
0
20
Value
40
60
Interpreting Histograms
Mean?
Median?
Mode?
Standard
deviation?
• Variance?
• Skew?
= 9.2
= 6.5
=3
= 8.3
40
Frequency
•
•
•
•
20
(which way does the
“tail” point?)
• Shape?
0
0
20
Value
40
60
An alternative:
Boxplots
Frequency
40
20
0
0
20
40
Value
Boxplots also
summarize a lot
of information...
Within each sample:
“Outliers”
Compared across samples:
Weight (kg)
75% percentile
Median
25% percentile
6
5
4
3
2
1
SC
Ba
Se
Es
Island
Da
Is
Fe
Normality
The Normal Distribution
aka “Gaussian” distribution
• Occurs frequently in nature
• Especially for measures that
are based on sums, such as:
– sample means
– body weight
– “error”
(aka “the Central Limit Theorem”)
• Many statistics are based on
the assumption of normality
– You must make sure your data
are normal, or try something
else!
Sample normal data:
Histogram + theoretical distribution
(i.e. sample vs. population)
Properties of the Normal Distribution
• Symmetric
Mean = Median = Mode
• Theoretical percentiles can be computed exactly
~68% of data are within 1 standard deviation of the mean
>99% within 3 s.d.
“skinny tails”
>99%
~95%
~68%
Amazing!
Handy!
Important!
What if my data aren’t Normal?
• It’s OK!
• Although lots of data are Gaussian (because of the CLT),
many simply aren’t.
– Example: Fire return intervals
• Solutions:
– Transform data to make it
normal (e.g. take logs)
– Use a test that doesn’t
assume normal data
Time between fires (yr)
• Don’t worry, there are plenty
• Especially these days...
• Many stats work OK as long as data are “reasonably” normal
Break!
Inferential Statistics:
Introduction
Inference: the process by which we draw
conclusions about an unknown based on
evidence or prior experience.
In statistics: make conclusions about a
population based on samples taken from
that population.
Important: Your sample must reflect the
population you’re interested in, otherwise
your conclusions will be misleading!
Statistical Hypotheses
• Should be related to a scientific hypothesis!
• Very often presented in pairs:
– Null Hypothesis (H0):
Usually the “boring” hypothesis of “no difference”
– Alternative Hypothesis (HA)
Usually the interesting hypothesis of “there is an effect”
• Statistical tests attempt to (mathematically)
reject the null hypothesis
Significance
• Your sample will never match H0 perfectly,
even when H0 is in fact true
• The question is whether your sample is
different enough from the expectation under
H0 to be considered significant
• If your test finds a significant difference, then
you reject H0.
p-Values Measure Significance
The p-value of a test is the probability of observing data
at least as extreme as your sample, assuming H0 is true
• If p is very small, it is unlikely that H0 is true
(in other words, if H0 were true, your observed sample would be unlikely)
• How small does p have to be?
– It’s up to you (depends on question)
– 0.05 is a common cutoff
• If p<0.05, then there is less than 5% chance that you would observe
your sample if the null hypothesis was true.
‘Proof’ in statistics
• Failing to reject (i.e. “accepting”) H0 does not
prove that H0 is true!
• And accepting HA doesn’t prove that HA is true
either!
Why?
• Statistical inference tries to draw conclusions
about the population from a small sample
– By chance, the samples may be misleading
– Example: if you always accept H0 at p=0.05, then
1 in 20 times you will be wrong!
Errors in Hypothesis Testing
• Type I Error: Reject H0 when H0 is actually true
– i.e. You find a difference when really there is none
– The probability of Type I error is called the
“significance level” of a test, and denoted α
• Type II Error: Accept H0 when H0 is actually false
– i.e. There really is a difference, but you conclude
there is none
– The probability of Type II error is denoted β
(and [1 – β is called the “power” of the test)
Assumptions of inferential statistics
• All inferential tests are based on assumptions
– If your data cannot meet the assumptions, the
test results are invalid!
• In particular:
– Inferential tests assume random sampling
– Many tests assume the data fit a theoretical
distribution (often normal)
• These are “parametric tests”
• Luckily, there are non-parametric alternatives
Inferential Statistics:
Methods
Student’s t-Test
Student’s t-test
• Several versions, all using inference on a
sample to test whether the true population
mean (µ) is different from __________
– The one-sample version tests whether the
population mean equals a specified value, e.g.
• H0: µ = 0
– The two-sample version tests whether the means
of two populations are equal
• H0: µ1 = µ2
t-Test Methods
• Compute t
– For one-sample test:
Remember:
x-m
t=
SE
s
SE =
√n
– For two-sample test:
t=
x 1 – x2
Sx1-x2
Sx1-x2 =

S1 2
n1
+
S2 2
n2
t-Test Methods
• Once you have t, use a look-up
table or computer to check for
significance
• Significance depends on
“degrees of freedom” (basically,
sample size)
• Bigger difference in means and
bigger sample size both improve
ability to reject H0
How does the t-Test work?
• Statistical magic!
• Student figured out the “tdistribution” shown to the
left.
– Given a sample mean and df, we
can see where the mean is likely
to be.
• If the null-hypothesis
mean is very unlikely to
fall under the curve, we
reject H0
Reject H0 if your t is in one
of the red regions!
a = 0.05
Try it!
Work through the Excel exercises for
one- and two-sample t-tests now
ANOVA
ANOVA: ANalysis Of VAriance
• Tests for significant effect of 1 or more factors
• Each factor may have 2 or more levels
• Can also test for interactions between factors
• For just 1 factor with 2 levels, ANOVA = t-test
– So why can’t we just do lots of t-tests for more
complicated experiments?
• Example: We study tree growth rates on clay vs.
sand vs. loam soil (10 trees each)
– How many factors? How many levels for each factor?
ANOVA: Concept
• Despite the name, ANOVA really looks for
difference in means between groups (factors &
levels)
• To do so, we partition the variability in our data
into:
– (1) The variability that can be explained by factors
– (2) The leftover unexplained variability (error or
residual variability)
Total variability = Variability due to factors + error
(we only have to calculate two of these values)
ANOVA example continued
Here are the raw data:
Square = clay soil
Diamond = sand soil
Triangle = loam soil
‘replicate’ = plot
‘y’ = growth
ANOVA example continued
• First find total variability
using Sum of Squares
– Find overall mean
(horizontal line)
– Each “square” is the
distance from one data
point to the mean,
squared
– Total Sum of Squares
(SST) is the sum of all the
squared deviations
ANOVA example continued
• Now measure variability
unexplained by factor of
interest (soil)
– Find means for each level
– Error Variability (SSE) is
the sum of all squared
deviations from these
level means
– Which is greater—SSE or
SST?
The remaining variability is due to soil factor (say, SSF). It’s easy to compute, since
SST = SSE + SSF
ANOVA example continued
• Next, we calculate degrees of freedom (df )
– df is based mainly on sample size
– Every time you estimate a parameter from your
data, you lose one df
• Example: Since we computed the mean of our 30
observations, we only need to know 29 of the values
now to determine the last one!
• For our example we have:
dfSST = 30 – 1 = 29
dfSSE = (10 – 1) + (10 – 1) + (10 – 1) = 27
dfSSF = dfSST – dfSSE = 2
ANOVA example continued
• From SS and df, we compute Mean Square (MS) variability
• Finally (!) we test whether the variability explained by our factor
is significant, relative to the remaining variability
– The ratio MSsoil/MSerror is F
– By statistical magic, we can look up the probability of observing such a
large F just by chance.
In other words, we find the p-value associated with H0: Soil has no effect on growth
Source
Soil
Error
Total
SS
99.2
315.5
414.7
df MS
F
2
49.6
4.24
27
11.7
29
p
0.025
What do
we
conclude?
• We can then go back and see which groups differ (e.g. by t-test)
Try it!
Work through the Excel exercise for
ANOVA now
Chi-square (χ2) Test
• In biology, it’s common to measure frequency (or
count) data
• χ2 is used to measure the deviation of observed
frequencies from an expected or theoretical
distribution:
c2 =
S
(O – E)2
E
Where: O is the observed frequency (# of events, etc.)
E is the expected frequency under H0
Should χ2 be big or small to reject H0?
χ2 Example
• Again we’ll also need to know degrees of
freedom
• For the χ2 test,
df = number of groups – 1
• Then (again by statistical magic), we can look
up how big χ2 needs to be (the critical value)
to reject H0 at a given significance level
χ2 Example
Imagine we conduct an experiment to determine the food
preference of rodents, with the following results:
Food
# Eaten
Tuna
Peanut butter
Fresh fish
Cheese
31
69
35
65
n=
200
A reasonable expectation (our null hypothesis) is:
H0 = Rodents like all foods equally well
Thus, under H0 our expected frequency for each food is:
200 / 4 = 50
χ2 Example
First, we draw up a contingency table:
tuna
PB
fish
cheese
Observed
31
69
35
64
200
Expected
50
50
50
50
200
Then we compute χ2:
c =
2
(31 - 50)2
50
+
(69 - 50)2
50
(35 - 50)2
+ 50
+
(65 - 50)2
50
= 22.0
χ2 Example
In our example:
c2 = 22.0
and
df = 4 – 1 = 3
The critical value for df = 3, α = 0.05 is c20.05,3 = 7.815
Since our χ2 is greater than χ2critical, we reject H0. Our
results differ significantly from the expectation under
H0, suggesting that there actually is a food preference
in the studied rodents.
Try it!
Work through the Excel exercise for
the Chi-square test now
Correlation
• Correlation measures the strength of the
relationship between two variables
– When X gets larger, does Y consistently get larger
(or smaller)?
• Often measured with Pearson’s correlation
coefficient
– Usually just called “correlation coefficient”
– Almost always represented with the letter
r
Correlation
Computing Pearson’s correlation coefficient:
r=
xy
S
x
S
2
y
S
2
≈
Amount that X and Y
vary together
Total amount of
variability in X and Y
-1 ≥ r ≥ 1
Correlation Examples
Correlation Cautions
• Correlation does not imply causality!
(note: doesn’t matter which data are X vs. Y)
• r can be misleading
– it implies nothing about slope
– it is blind to outliers and obvious
nonlinear relationships
Same r in each
panel!
Try it!
Work through the Excel exercise for
correlations now
Regression
• Unlike correlation, regression does imply a
functional relationship between variables
– The dependent variable is a function of the
independent variable(s)
• In regression, you propose an algebraic model,
then find the “best fit” to your data
– Most commonly, the model is a simple line (“Y is a
linear function of X”)
Regression
There are many possible relationships between two variables...
Linear
Exponential
Quadratic
Hyperbolic
Logistic
Trigonometric
Regression
• We’ll focus on simple linear regression of one
dependent (Y) and one independent (X) variable
• The model is:
Y = a + bX + 
Y = values of the dependent variable
X = values of the independent variable
a, b = “regression coefficients” (what we want to find)
 = residual or error
Potential Regression Outcomes
Positive
Relationship
b>0
Negative
Relationship
b<0
No
Relationship
b=0
Y
X
X
X
In regression we always plot the independent variable on the x-axis
How do we “fit” a Regression?
• Most common method is “least-squares”
– Find a and b to minimize the (squared) distances of data
(X4, Y4) .
points from the regression line

Y –Y = 
.
^
(X3, Y3)
5
.
(X5, Y5)
Y
(X2, Y2)
.
.
(X6, Y6)
. (X1, Y1)
X
5
5
How do we “fit” a Regression?
Find individual residuals (ε):
^
Yi – Yi =
i
Observed value
Residual
Predicted value
(from regression line)
Then the sum of all
(squared) residuals is:
(Y
–
S i
^
2
Yi)
A computer or clever mathematician can find the a and b that
minimize this expression (producing the “best fit” line)
Regression Example
Altitude (m)
0
50
190
305
456
501
615
700
825
Temperature (C)
25.0
24.1
23.5
21.2
20.6
20.0
18.5
17.2
17.0
Which is the independent variable?
Regression Example
From the data on the previous slides, we fit a regression line and
obtain these parameter estimates:
a = 24.88
b = 0.0101
Thus, our best-fit line has the equation:
temperature = 24.88 – 0.0101 x altitude
Does this work if we change units?
Regression: How good is the fit?
.
..
.
.
.
.
.
..
.
Perfect fit - all the
points on the line
..
. .
.
.
. .
. .
.
.
.
..
.
.
.
.
Good fit
OK fit (I’ll take it!)
Regression: How good is the fit?
• Formally, we often measure the fit with the
coefficient of determination, R2.
• R2 is the proportion of variation in Y “explained” by
the regression
– Values range from 0 to 1
– 0 indicates no relationship, 1 indicates perfect
relationship
Note: In simple linear regression, yes, R2 is actually the same as
Pearson’s correlation coefficient (r) squared. But this is just a special
case—regressions get much more complicated. Don’t get in the habit
of confusing the two statistics!
Regression: Is the fit significant?
• We usually ask whether R2 is significant by testing
the null hypothesis:
H0: b = 0 (the line may as well be flat)
against the alternative:
HA: b ≠ 0 (the best fit isn’t a flat line)
• Luckily (statistical magic), this can be tested with
the t-test!
– Depends on degrees of freedom
(increase sample size to improve significance)
– For our purposes, statistics software will do this for you
Regression Reservations
• Again, regression does imply causality (unlike
correlation), but importantly, it still does not
test a causal biological relationship
– Some other variable might affect both X and Y
– You could even have the relationship backwards!
• Be careful extrapolating regression lines:
– beyond the original data range, or
– to other populations
Try it!
Work through the Excel exercise for
simple linear regression now